IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 59, NO.

4, MAY 2010

2057

Analysis of Flip Ambiguities for Robust

Sensor Network Localization
Anushiya A. Kannan, Member, IEEE, Bars Fidan, Member, IEEE, and Guoqiang Mao, Senior Member, IEEE

AbstractErroneous local geometric realizations in some

parts of a network due to the sensitivity to certain distancemeasurement errors with respect to some neighboring sensor locations is a major problem in wireless sensor-network localization,
which may, in turn, affect the reliability of the localization of the
whole or a major portion of the sensor network. This phenomenon
is well described using the notion of flip ambiguity in rigid
graph theory. In this paper, we present a formal geometric analysis
of flip-ambiguity problems in planar sensor networks via quantification of the likelihood of flip ambiguities in arbitrary sensor
neighborhood geometries. Based on this analysis, we establish a robustness criterion to detect flip ambiguities in such neighborhood
geometries. In addition to the analysis, the established robustness criterion is embedded in localization algorithms to enhance
the reliability of the produced location estimates by eliminating
neighborhoods with flip ambiguities from being included in the
localization process.
Index TermsFlip ambiguities in WSN localization, robust
WSN localization.

I. I NTRODUCTION

ECENT advances in integrated circuit design, embedded

systems, and wireless communication technology have
enabled smart tiny sensors to be deployed in large numbers.
Most of these sensor nodes will be deployed at positions that
may not be predetermined due to constraints on the implementation environment and/or the cost of deployment at known
locations. The information gathered by such sensor nodes, in
general, will be useless without determining the locations of
these nodes. This makes self-localization capabilities a highly
desirable characteristic of sensor networks [1][4]. Sensor
network-localization algorithms use intersensor measurements,
which can be in the form of distance measurements, bearing
measurements, time-difference-of-arrival measurements, etc.,
and the a priori known locations of some specific sensors,

Manuscript received June 22, 2009; revised October 12, 2009. First published February 8, 2010; current version published May 14, 2010. This work
was supported by the National Information and Communications Technologies
(ICT) Australia, which is funded by the Australian Government as represented
by the Department of Broadband, Communications, and the Digital Economy
and the Australian Research Council through the ICT Centre of Excellence
Program. The review of this paper was coordinated by Dr. P. Lin.
A. A. Kannan is with the University of Sydney, Sydney, NSW 2006,
Australia, and also with the Biomedical Systems Laboratory, University of
New South Wales, Sydney, NSW 2052, Australia (e-mail: Anushiya.Kannan@
nicta.com.au).
B. Fidan is with the Department of Mechanical and Mechatronics Engineering, University of Waterloo, Waterloo, ON N2L 3G1, Canada (e-mail: fidan@
uwaterloo.ca).
G. Mao is with the School of Electrical and Information Engineering,
University of Sydney, Sydney, NSW 2006, Australia (e-mail: Guoqiang.Mao@
nicta.com.au).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TVT.2010.2040850

which can be usually obtained via the Global Positioning

System or by deploying sensors at known locations, to estimate
the locations of all sensors in the network. This paper focuses
on distance-measurement-based localization algorithms.
Based on the approach of processing the distance measurements, distance-based localization algorithms can be categorized into two main classes [5]: centralized algorithms and
distributed algorithms. Centralized algorithms [5][8] use a
single central processor to collect all distance measurements
between neighboring nodes and produce a map of the entire
sensor network, while distributed algorithms [5], [9][11] rely
on the self-localization of each sensor node in a sensor network
using the distance measurements they collect from one-hop
and multihop neighbors and the a priori known or estimated
locations of these neighbors.
A fundamental problem in distance-based sensor network
localization is whether a given sensor network is uniquely
localizable or not, i.e., whether the sensor network geometry
corresponding to known locations of some specific sensors and
a given set of intersensor distance measurements is unique or
not. A particular framework that is useful for analyzing this
problem is rigid graph theory [12][15]. In this framework, the
sensor network to be localized is modeled by a graph called
the underlying graph of the sensor network, where the vertices
of the graph represent sensor nodes, and all the vertex pairs
corresponding to the sensor node pairs with known distance
measurements are connected with edges.
If the actual geometries of some parts of the sensor network
are sensitive to distance measurement errors, it may cause
erroneous local geometric realizations in those parts of the
network.1 This is a major problem affecting the unique realizability of the distance-based localization [4], [6], [7]. One
such phenomenon is flip ambiguity, which can prevent unique
localization of the sensor network [14][17] (in the sense that
it differs from other such representations at most by translation,
rotation, or reflection).
In flip ambiguities, a sensor node (or sensor nodes) with a set
of neighbors that are almost collinear may lead to the possibility
of the neighbors forming a mirror through which the sensor
node (or the sensor nodes) can be reflected, thereby causing a
large localization error. Moreover, when the localization algorithm proceeds, it is likely to continue degrading the location
estimates in the subsequent iterations that use the location estimate of the flipped sensor nodes. This impact can propagate in
an avalanche fashion for several iterations, affecting, in turn, the
location estimates of either the entire network or a large portion
1 Although similar problems may also exist in localization tasks based on
other types of measurement techniques, we focus only on distance-based
localization.

0018-9545/$26.00 2010 IEEE

2058

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 59, NO. 4, MAY 2010

Fig. 1. Avalanche error propagation of flip ambiguity seen in a centralized

localization algorithm presented in [8]. The anchor nodes marked as 1, 2, and
3 are placed almost collinearly and cause a large number of flipped realizations
in the location estimates of sensor nodes shown within the rectangle.

of it. This behavior of avalanche error propagation can be

noticed in many localization algorithms in the literature. Fig. 1
shows one such avalanche error propagation in a centralized
algorithm [8]. In the bottom part of Fig. 1, the three anchors
noted as anchors 1, 2, and 3 are placed almost collinearly.
This causes a large number of sensor nodes (shown within the
rectangle in Fig. 1), which may not necessarily be the onehop neighbors of these three anchors, to be flipped to incorrect
positions, thereby resulting in large location-estimation errors
in the system. A detailed description of flip ambiguities and
the resulting avalanche error-propagation effects are given in
Sections III and IV.
An empirical study on the frequency of flip-ambiguity occurrences is presented in [18]. This study shows that with an
increasing number of neighbors of a location-unknown sensor
node, the frequency of flipped realizations reduces (but never
gets to zero). Moreover, at the initial stages of any localization
algorithm, many neighbors will not know their locations, and
locations are estimated with a smaller number of neighbors,
causing more frequent occurrences of flip ambiguities. When
the localization algorithm proceeds, it is likely for the number
of neighbors to grow, reducing the frequency of flip ambiguities. However, if the set of neighbors contains any flipped
realized sensor node, it can cause the flip-ambiguity effect to
propagate in the rest of the localization process. This may, in
turn, affect the localization of the whole or a major part of the
sensor network, as explained in Section IV.
Flip ambiguities do not necessarily occur in every sensor
network, but when they occur, they can significantly degrade
the localization of the whole network (as shown in Fig. 1). It is
therefore important to identify possible flip ambiguities in the
location estimates and take proper action to prevent them from
being used in the localization process.
This paper proposes a robustness criterion to identify the
likelihood of flip ambiguities in arbitrary sensor neighborhoods,
which is applied on all neighborhoods to identify the sensor
nodes contributing to flip ambiguity and eliminate them from
the neighborhoods used in the localization process. Thus, with
the help of the robustness criterion, the localization algorithm
selects the robust neighborhoods containing only the sensor
nodes that do not contribute to flip ambiguities. Even though the
selection of robust neighborhoods will mitigate flip ambiguity
in both centralized and distributed localization algorithms, this
paper focuses only on the performance improvement in distributed localization.

The rest of this paper is organized as follows. Section II summarizes the related work in the literature. Section III presents
a brief overview of a theoretical framework of unique sensor
network localization and the associated ambiguities. Section IV
explains the effects of flip ambiguities in localization algorithms using simple examples and how error propagates following the occurrence of a flipped realization. Section V analyzes
the effects of neighborhood geometries on flip ambiguities
and formulates the problem of identifying flip ambiguities in
sensor neighborhoods. Section VI proposes a robustness criterion to identify possible neighborhoods with a high probability
of having flip ambiguities. Section VII presents a numerical
analysis of the effectiveness of the derived robustness criterion
in the removal of flipped realizations. Section VIII explains
a localization algorithm using our robustness criterion and
presents the simulation results demonstrating the performance
enhancement. Section IX presents a more application-oriented
version of the localization algorithm presented in Section VIII,
which increases the number of localized nodes with negligible
degradation in estimation errors. This paper is concluded in
Section X.
II. R ELATED W ORK
In the literature, the flip-ambiguity problem has been approached from different perspectives to eliminate their effect
on the location estimation of the sensor nodes. The main aim of
these methods is to estimate the locations of the sensor nodes
with minimum error.
The notions of Voronoi diagrams and Delaunay graphs are
used in [19] and [20] to glue the Delaunay triangles in an
incremental fashion to achieve unique localization of boundary
or landmark sensors. These sensors are then used as anchors
to localize the rest of the network. Such a method requires the
boundary/landmark sensors to be sufficiently dense to maintain
the global rigidity of the Delaunay graph. The underlying graph
structure of the sensor network is explored in [14], [17], and
[19] to obtain unique localization. The algorithms in [14] and
[17] maintained global rigidity, which is a sufficient condition
for unique localization, by having a denser network. The concept of global rigidity is defined later in Section III-A.
Designing an efficient distributed algorithm for global rigidity, however, is nontrivial, as neither connectivity nor rigidity
can be tested locally by nature [21]. In practice, trilateration
[15], [22][26] is widely used to localize the unknown sensor
nodes, as it is fully distributed and easily implementable. The
study in [21] pointed out that wheel graphs are globally rigid
and that trilateration is a special case of wheel graphs.
Global rigidity is, however, only a sufficient condition for the
unique localization of a sensor network. A priori information
may compensate the need for global rigidity in some graph
structure [4]. Moreover, the graph realization theory does not
account for measurement errors affecting intersensor measurements in real applications [25]. For sparse networks with noisy
measurements, the algorithms in [27] and [28] recorded all
possible location estimates of each sensor and eliminate the
incompatible location estimates whenever possible, which, in
the worst case, can result in an exponential space requirement.
Another approach for mitigating flip ambiguity in noisy
environments is to identify possible flip ambiguities in location estimates and take the necessary action to eliminate such

KANNAN et al.: ANALYSIS OF FLIP AMBIGUITIES FOR ROBUST SENSOR NETWORK LOCALIZATION

flip ambiguities from the localization process [15], [25], [26].

The studies in [15] and [25] use the trilateration-based sensor
quadruple, which is a four-sensor neighborhood structure, to
derive robust criteria to identify possible flip ambiguity in
location estimates of sensor nodes. Only the sensor quadruples
that are identified as not susceptible to flip ambiguities by the
robustness criterion are tagged as robust neighborhoods and are
used in the localization process. The study in [26] provided a
formal geometric analysis of the flip-ambiguity problem using
similar notions as those in [15] and [25]. It developed a generic
formal method for quantifying the likelihood of flip ambiguities
for arbitrary sensor neighborhood geometries.
The analyses in [15], [25], and [26] have been conducted
under the assumption of unknown sensor nodes having access
to intersensor distance measurements from all three of its
neighbors, where only one distance measurement is considered
to be in error, while the other two are considered accurate.
Our paper also considers completely connected sensor
quadruples in a planar sensor network and establishes a robustness criterion to identify possible flip ambiguities. In contrast
to [15], [25], and [26], which consider unique localizability
with only one erroneous distance measurement while assuming
all the other distance measurements are accurate, our paper
accommodates the errors in one or more intersensor distance
measurements. Furthermore, unlike the work in [15], [25], and
[26], which restricted the localization to trilateration using the
sensor quadruples, our work uses a novel way to select all
neighbors that do not contribute to flip ambiguities (which is
explained in Section VIII) and to use them in the localization
algorithm with a multilateration scheme.

III. G RAPH -T HEORETICAL F RAMEWORK FOR

S ENSOR N ETWORK L OCALIZATION
A. Modeling for Unique Sensor Network Localization
In a graph-theoretical framework, a sensor network can be
represented by a graph G = (V, E) with a vertex set V and an
edge set E, where each vertex X V is uniquely associated
with a sensor node sX in the network, and each edge (X, Y )
E uniquely corresponds to a sensor node pair (sX , sY ) for
which the intersensor distance is known [12][17], [29]. The
graph G(V, E) is called the underlying graph of the network.
The planar location information about the sensor nodes corresponds to a 2-D representation of the representative graph,
which is a mapping p : V R2 , assigning a location in R2 to
each vertex in V . Given a graph G = (V, E) and a representation of it, the pair (G, p) is called a framework.
A particular graph property associated with the unique realizability of sensor networks is global rigidity [4], [12], [13],
[29]. A framework (G, p) is globally rigid if every framework (G, p1 ) satisfying p(X) p(Y ) = p1 (X) p1 (Y )
for any vertex pair X, Y V , which are connected by an edge
in E, also satisfies the same equality for any other vertex pairs
that are not connected by an edge. A relaxed form of global
rigidity is rigidity: a framework (G, p) is rigid if there exists a
sufficiently small positive constant  such that every framework
(G, p1 ) satisfying 1) p(X) p1 (X) <  for all X V and
2) p(X) p(Y ) = p1 (X) p1 (Y ) for any vertex pair
X, Y V , which are connected by an edge in E, also satisfies

2059

Fig. 2. Discontinuous flex ambiguity. By removing the edge AD, flexing the
edges AB, BC, and AE, and reinserting the edge AD, a different realization
can be obtained while satisfying the same distance constraints.

Fig. 3. (a) Flip ambiguity in a rigid but not globally rigid underlying graph. By
reflecting D across the line connecting sensor nodes A and B, we obtain a new
satisfying the same distance constraints. (b) Flip ambiguity in a
realization D
globally rigid underlying graph with a near-collinear neighborhood. Depending
on the sensitivity of the near-collinear neighborhood ABC to the intersensor

distance measurement errors, D can be estimated at a flipped location D

satisfying approximately the same distance constraints.

the equality in condition 2 for any other vertex pairs that are not
connected by an edge.
B. Ambiguities in Sensor Network Localization
If a framework (G, p) is rigid but not globally rigid, there
exist two types of discontinuous deformations that can prevent
a representation of G from being consistent with p, i.e., a
representation (G, p1 ) satisfying p(X) p(Y ) = p1 (X)
p1 (Y ) for any vertex pair X, Y V , which are connected
by an edge in E, from being unique (in the sense that it
differs from other such representations at most by translation,
rotation, or reflection) [15], [16]: flip and discontinuous flex
ambiguities.
Discontinuous flex ambiguities arise when the removal of an
edge or a set of edges allows the remaining part of the graph to
be flexed continuously to a different realization while satisfying
the same distance constraints when the removed edge(s) is (are)
reinserted (see Fig. 2).
When an underlying graph is rigid but not globally rigid, a
vertex (sensor node) of the underlying graph may have a set
of neighbors forming a mirror across which the vertex (sensor
node) can be reflected. Such a phenomenon is called a flip
ambiguity, which is depicted in an example in Fig. 3(a).
However, even in a sensor network with a globally rigid
underlying graph, a sensor node can still be reflected across
its near-collinear neighbors depending on the sensitivity of
the near-collinear neighborhood to the corresponding distance
measurement errors [4]. Fig. 3(b) depicts an example of such
situations. Thus, erroneous measured distances may cause flip
ambiguities on near-collinear neighborhoods, depending on
the sensitivity of those near-collinear neighborhoods to such
measurement errors.
IV. E RROR -P ROPAGATION E FFECTS OF F LIP A MBIGUITIES
This section, using a simple example, describes how flipambiguity effects can propagate in a sequential or incremental

2060

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 59, NO. 4, MAY 2010

each other, i.e., the distance between any pair is measurable.

In the localization process, a robust neighbor set RNX NX
is formed for every unknown sensor node X by selecting all
members Y NX that do not contribute to flip ambiguity,
using the robustness criterion. Hence, the robustness criterion
helps the localization algorithm to use information from all
known neighbors that do not contribute to flip ambiguity.
V. Q UANTITATIVE A NALYSIS OF F LIP A MBIGUITIES
Fig. 4. Avalanche error propagation of a single flip ambiguity. Initially, the
locations of A, B, C, P , and Q are known, and D, E, F , G, H, and I are
unknown. At first, D and E uses the locations of near-collinear neighbors A,
and E
[satisfying (38)] due to flip
B, and C and localize themselves at D
ambiguity. Then, F uses the locations of neighbors A, C, and D and localizes
Subsequently, G, H, and I will get localized at G,
H,
and I,

itself at E.
respectively. Here, a flip ambiguity in D (and E) caused nodes F , G, H, and I
to have large estimation errors.

localization [22], [23], [30][32]. A similar explanation of

flip-ambiguity effects in a cluster localization followed by
stitching/patching [15], [24], [33][35] can be found in [18].
For convenience, we call a sensor with a priori known location
or whose location has already been estimated as an anchor or a
known sensor node and a sensor with unknown location as an
unknown sensor node.
In a sequential localization algorithm [22], [23], [30][32],
unknown sensor nodes are localized once they have three or
more known neighbors. As soon as a sensor node is localized,
it will be elevated to have anchor status, increasing the number
of known neighbors of some unknown sensor nodes and may
make them eligible for localization. Sequential localization
continues until there are no unknown sensor nodes with at
least three known neighbors. Flip ambiguity in any location
estimate may cause an avalanche effect in the subsequent
location estimates, thereby causing large estimation errors.
Such a phenomenon is demonstrated in Fig. 4 via a simple
example. We call the aforementioned kind of error propagation
as avalanche error propagation, where the estimation error
multiplication allows very large estimation errors to propagate
in the network starting with the flipped realization of a single
node. The error-propagation effects, where the location estimation error increases with the sequence of location estimates,
become apparent in sequential localization algorithms when the
number of anchors is small.
From the aforementioned illustration, we can see that when
a flip ambiguity occurs, it may drastically corrupt the sensor
network localization or, in the extreme case, may even make
the whole network localization collapse. Hence, it is necessary
to identify the possible occurrence of flip ambiguity in any
neighborhood to avoid large localization errors. Finally, we
also want to point out that such avalanche error-propagation
phenomenon also exists in centralized localization algorithms,
as shown in Fig. 1 and [8].
The robustness criterion proposed in Section VI can be used
by localization algorithms to make a binary decision of whether
to accept or to reject a neighbor node being considered as a
member of a neighborhood in a localization process. To keep
the analysis simple, the robustness criterion is formed only
by considering fully connected quadrilaterals (FCQs), which
are quadruples of sensor nodes, all of which are neighbors of

This section analyzes the conditions under which a flip ambiguity is likely to occur in an arbitrary FCQ of a given sensor
network. The analysis assumes a disc transmission model of
radius R around each sensor node X.
Assumption 1: Two sensor nodes X and Y are said to be
neighbors of each other and are able to measure their intersensor distance to each other if and only if the true distance
|XY | R. In accordance with this disc transmission model,
any intersensor distance measurement dXY > R is assumed to
be truncated to R.
Furthermore, as justified in [36], this analysis uses a
Gaussian distribution with zero mean and a standard deviation
of as the noise model for the distance measurement errors.
Thus, it is possible to state that the absolute value of the distance
measurement error is smaller than a threshold  > 0 with a
certain probability. To keep this probability high,  can be
chosen as 3, such that the probability of the absolute value
of the distance measurement error being less than  is 99%.
Based on this observation, we consider a truncated version of
the Gaussian distribution and state the following assumption.
Assumption 2: For every neighbor node pair X, Y NX ,
the difference between the true distance |XY | and the measured
distance dXY is within a known value  > 0, i.e.,


|XY | dXY  .
(1)
In a localization problem, once the positions of three sensor
nodes in an FCQ are precisely known and the measurements of
the distances of the fourth sensor node from each of these three
sensor nodes are available, the fourth sensor node is localizable.
Consider an ordered FCQ ABCD as a quadruple of sensor
nodes A, B, C, and D, where the locations of A, B, and C are
known together with intersensor distance measurements dAD ,
dBD , and dCD . The localization task focuses on the particular
problem of estimating the location of sensor node D using
the given fixed positions of neighbor nodes A, B, and C and
intersensor measurements dAD , dBD , and dCD .
Following the notation in parallel with [15] and [26], an FCQ
is called (, S )-robust (with respect to distance measurement
errors) if the fourth sensor node (called sensor node D) is
uniquely localizable with a predefined accuracy level S > 0
when the corresponding absolute measurement errors are less
than a predefined threshold  > 0.
The analysis is started with the focus on the following
generic localization method.
Step 1) Using dAD and dBD only, find the two possible
and D
 for sensor node D as intersection
locations D
points of the circles C(A, dAD ) with center A and
radius dAD and C(B, dBD ) with center B and radius
and D
 are symmetrical
dBD . These two points D
with respect to AB.

KANNAN et al.: ANALYSIS OF FLIP AMBIGUITIES FOR ROBUST SENSOR NETWORK LOCALIZATION

2061

cause a substantial error in the estimated location of a sensor

is called a negligible flip ambiguity. Hence, a flip ambiguity is
 S2 symmetrical
S1 and a D
substantial if there exist a D

to D with respect to AB such that


1 
 |  when |D
D
 | > S .
(6)
|C D| |C D
2
Fig. 5. Two regions S1 and S2 for the possible location-estimation pair
D
 ) defined by the intersection of the two rings R1 and R2 represented
(D,
by (4) and (5).

Step 2) Using the third distance measurement dCD , decide

and D
 to choose as the locaon which of D
tion estimate of D that better matches the distance
constraints.

Hence, the criterion to judge the (, S )-robustness of an FCQ

can be stated as

1 
 | > 
D| |C D
min
|C

DS
2
1
 symmetrical to D
w.r.t. AB
D

D
 | > S
when |D

(7)

Remark 1: Without loss of generality, the aforementioned

processing order of the intersensor measurements (dAD , dBD ,
dCD ) is chosen, but the analysis below can easily be adapted to
any other order of distance measurements.
This generic localization method does not consider the
distance measurement errors in dAD and dBD . Using
Assumption 2, for a given , the erroneous measured distances
dAD , dBD , and dCD can be bounded using the true distances
|AD|, |BD|, and |CD| as

for a given neighborhood of A, B, C and measured distances

dAD and dBD . Note that the three distance measurement errors
in dAD , dBD , and dCD are simultaneously considered in (7).
Distance measurement errors in dAD and dBD determine the

regions S1 and S2 , and the minimization of (1/2)(C D|


|C D ) over these regions means that (7) considers the errors

in both dAD and dBD . The error in dCD is explicitly considered
in the inequality in (7).

dXD [ |XD| , |XD| +  ] ,

B. Analysis of Flip Ambiguities Using Any Pair of Distance

Measurements dAD , dBD , and dCD

X {A, B, C}.

(2)

Equivalently
|XD| [ dXD , dXD +  ],

X {A, B, C}.

(3)

A. Analysis of Flip Ambiguities Using a Particular Distance

Measurement Pair dAD and dBD
In this section, the criterion for flip ambiguity is analyzed
using the erroneous distance measurements of sensor node D
from neighbors A and B first. In the presence of distance

measurement errors in dAD and dBD , the possible locations D

 of sensor node D are inside the two regions S1 and S2
and D
(see Fig. 5), which are the intersection of the two rings R1 and
R2 defined as

R1 = (x, y) | ( dAD  )2 (x xA )2 + (y yA )2

(4)
( dAD +  )2

R2 = (x, y) | ( dBD  )2 (x xB )2 + (y yB )2

(5)
( dBD +  )2
where (xA , yA ) and (xB , yB ) are the known coordinates of
and D
,
sensor nodes A and B, respectively. Here, the points D
as well as the regions S1 and S2 , are symmetrical with respect
to AB.
The generic task of sensor localization is estimating the
location of each sensor node such that the magnitude of the
corresponding estimation error is less than or equal to a predefined accuracy level S > 0 when the corresponding absolute
measurement errors are less than a predefined threshold  > 0.
D
 | < S
Hence, an error caused by a flip ambiguity where |D
is not substantial. In this paper, a flip ambiguity that does not

The aforementioned analysis analyzed the criterion for

flip ambiguities using the distance measurements of sensor
node D from neighbor nodes A and B first to form the
regions S1 and S2 . If instead, the distance measurements of
sensor node D from neighbor nodes A and C are considered
and D
 of sensor node D will
first, the two possible locations D
AC
be inside two different regions S1 and S2AC symmetrical with
respect to AC, leading to a different robustness criterion

1 
 | > 
D| |B D
min
|B
AC

DS
2
1
 symmetrical to D
w.r.t. AC
D

D
 | > S . (8)
when |D
Hence, the criterion to judge the (, S )-robustness of an
FCQ ABCD should satisfy all three permutations of the
sequence A, B, and C as stated as follows:

1 
 | > 
D| |Z D
min
|Z
XY

DS
2
1
 symmetrical to D
w.r.t. XY
D
X,Y,Z{A,B,C},X=Y =Z

D
 | > S . (9)
when |D

VI. I DENTIFICATION OF P OSSIBLE F LIP A MBIGUITIES

A. Identification of Flip Ambiguities Using a Particular
Distance Measurement Pair dAD and dBD
This section focuses on the minimum of the flip-ambiguity
|C D
 || and formulates a robustness criterion
indicator ||C D|
of FCQs ABCD while using the distance measurements of
sensor node D from neighbors A and B first. Without loss of

2062

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 59, NO. 4, MAY 2010

generality, let the coordinates of A and B lie on the x-axis

with the midpoint of AB at the origin. Then, (xB , 0) and
(xB , 0) are the coordinates of A and B, respectively, where
xB = |AB|/2. Let (xC , yC ) be the coordinate of sensor node
C. With these definitions, (4) and (5) can be rewritten as

R1 = (x, y) | ( dAD  )2 (x + xB )2 + y 2

(10)
( dAD +  )2

R2 = (x, y) | ( dBD  )2 (x xB )2 + y 2

(11)
( dBD +  )2 .
be in
Without loss of generality, let the location (x, y) of D

be in S2 , where S1 and S2
S1 and the location (x, y) of D
are defined using (10) and (11) as
S1 = {(x, y) | (x, y) (R1 R2 )} and y > 0

(12)

S2 = {(x, y) | (x, y) (R1 R2 )} and y > 0. (13)

The regions S1 and S2 are considered to be disjoint in the
initial analysis.
If an error function e is defined as



|C D
 |
(14)
e = |C D|
then (7) can be written as
min

DS
1
 symmetrical to D
w.r.t. AB
D

D
 | > S .
e2 > 4 2 when |D

(15)

From (12), (13), and Fig. 6, it is seen that when (xS , yS ),

(xQ , yQ ), (xR , yR ), and (xP , yP ) are the coordinates of the
intersection points S, Q, R, and P , respectively, then the range
and D
 , respectively,
of the coordinates (x, y) and (x, y) of D
satisfying (12) and (13), is bounded as follows:
xS x xQ

and

y P y yR .

(16)

and
To further analyze (15), let us write |C D|



|C D | in terms
 of the coordinates of C, D, and  D
2
2
= (xC x) + (yC y)
|=
as
|C D|
and
|C D

2
2
(xC x) + (yC + y) , giving (17), shown at the bottom of
the page.
From (17), we can see that for a fixed y, e2 is minimized
when |x xC | is maximized. Depending on the location of

sensor node C, the minimum of e2 will occur when either D

is on the left boundary of S1 (arcs RS and SP in Fig. 6) or D

is on the right boundary of S1 (arcs RQ and QP in Fig. 6). If
sensor node C is on the right side of the region, i.e., closer to

Fig. 6. Detailed illustration of the two regions S1 and S2 and the possible
D
 ) using the accurate locations of sensor nodes
location-estimation pair (D,
A and B, erroneous intersensor distance measurements dAD and dBD , and
the threshold value .

the right boundary of S1 than to the left boundary, the minimum

is on the left boundary of S1 ; otherwise, the
will occur when D
is on the right boundary of S1 .
minimum will occur when D
This can also be shown by examining the partial derivative
of e2 . If we take the partial derivative of e2 with respect to x

2
2
2
2
e
(xC x) + yC + y
= 4(xC x)
1.
x
2 + y 2 )2 4y 2 y 2
((xC x)2 + yC
C
(18)
Since
2

2
2 2
(xC x)2 + yC
+ y 2 4yC
y
 2
  2
2
= (xC x)4 + 2(xC x)2 yC
+ y 2 + yC
y2 > 0

2 + y 2 )2 4y 2 y 2 is always
((xC x)2 + yC
the value of
C
real. A closer look reveals that

2 + y 2 )2 4y 2 y 2 = 2|C D||C
D
 | (19)
((xC x)2 + yC
C
which is a product of two distances, and hence

2 + y 2 )2 4y 2 y 2 > 0.
((xC x)2 + yC
C
Moreover

2 + y 2 )2 4y 2 y 2
((xC x)2 + yC
C

2 + y 2 )2 = (x x)2 + y 2 + y 2 .
<
((xC x)2 + yC
C
C





2
2
2 + y 2 )2 4y 2 y 2
|C D
 | = 2 (xC x)2 + yC
e2 = |C D|
+ y 2 ((xC x)2 + yC
C
=
2 + y2 +
(xC x)2 + yC

2 2
y
8yC
2

2 + y 2 ) 4y 2 y 2
((xC x)2 + yC
C

(17)

KANNAN et al.: ANALYSIS OF FLIP AMBIGUITIES FOR ROBUST SENSOR NETWORK LOCALIZATION

It follows that
1< 
and hence

(xC x) +
2

2
yC

+y

2 + y 2 ) 4y 2 y 2
((xC x)2 + yC
C

0 < 

2
+ y2
(xC x)2 + yC

((xC

x)2

2
yC

2
y2 )

1
2 y2
4yC

enforcing that there is only one extremum for e2 , which is

at x = xC . It is also seen from (18) that e2 is an increasing
function (i.e., e2 /x > 0) when x < xC and a decreasing
function (i.e., e2 /x < 0) when x > xC . This leads to the
conclusion that, for any fixed value of y, the minimum of e2
S1 occurs at the boundaries of S1 where |xC x| is
when D
at its maximum. Rewriting the squared error term e2 as in (20),
shown at the bottom of the page, we get
2
lim e2 = 4yC

y+

From the aforementioned discussion, we can establish that

is on the lower
the minimum of e2 will only occur when D
boundary of S1 . When C is on the right side of S1 , the
is on the lower left boundary of S1
minimum occurs when D
(arc SP in Fig. 6); likewise, when C is on the left side of S1 ,
i.e., closer to the left boundary of S1 than to the right boundary,
is on the lower right boundary of
the minimum occurs when D
S1 (arc QP in Fig. 6). Without loss of generality let us focus
on the case when C is on the right side of S1 , and hence, the
minimum of e2 occurs when D is on the lower left boundary
is on the segment SP . In the following
of S1 , i.e., when D
and
paragraph, we shall further narrow down the range of D
2
is
establish that the minimum of e will only occur when D
either at point S or at point P .
Since arc SP is a part of the circle centered at (xB , 0) with
is on the segment SP , we can obtain
a radius dAD , when D
the following relationship between y and x:
y 2 = (dAD )2 (x + xB )2 .

lim e2 = 0 for any fixed value of x.

y0

For any fixed x, when y decreases, e2 also decreases. Given

the condition that yP y yR , it can be concluded that irrespective of the location of sensor node C, for any fixed x, the
is on the lower boundary
minimum of e2 will occur when D
(arcs SP and P Q in Fig. 6) of S1 . This conclusion can also be
obtained by examining the partial derivative of e2 .

((xC x)2 +yC2 )

y2

+1+

Substituting this relationship of x and y into e2 in (17), it has

the form shown in (21), shown at the bottom of the page.
Equations (22) and (24), shown at the bottom of the page,
show the first derivative of e2 with respect to x and the second
derivative of e2 with respect to x [obtained using (U/V )/x,
where U and V are as in (23), shown at the bottom of the page],

2
8yC

e2 =

2063

((xC x)2 +yC2 )

(20)

y4

2
2((xC x)2 yC
)
y2

+1




2
2 + y 2 )2 4y 2 y 2
e2 = 2 (xC x)2 + yC
+ y 2 ((xC x)2 + yC
C

2
= 2 (xC x)2 (xB + x)2 + yC
+ ( dAD  )2

(xC x)2 (xB + x)2 +

2
yC

+ ( dAD  )2


+

2 (x
4yC
B

+ x)2

2 (d
4yC
AD

 )2

(21)

+ (dAD ) (xC + xB )
+ x)


2
2
2
2
2
2
2
2
2
xC 2x(xC + xB ) xB + yC + ( dAD  )
+ 4yC (xB + x) 4yC ( dAD  )
(22)


2
2
U = x2C 2x(xC + xB ) x2B + yC
+ ( dAD  )2 (xC + xB ) 2yC
(xB + x)

 2

2 + (d
2 2 + 4y 2 (x + x)2 4y 2 ( d
2
xC 2x(xC + xB ) x2B + yC
(23)
V =
AD  )
B
AD  )
C
C



2
2
2
2
2
2 (e2 ) 8yC ( dAD  ) (xC + xB ) + yC
=
(24)
3
x2
(V ) 2
2

(e )
= 4 (xC + xB ) 
x

x2C

2x(xC + xB )

x2B

2
yC

2
2yC
(xB

2064

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 59, NO. 4, MAY 2010

respectively. From (17) and (21), it can easily be seen that V =

D
 | > 0. Hence, (24) shows that the second derivative
|C D||C
2
of e is always negative, i.e., e2 is a concave function of x when
is confined to the segment SP . The property of the concave
D
is on the
function readily leads to the conclusion that when D
2
segment SP , the minimum of e occurs only at the boundaries,
is either at point S or at point P .
i.e., when D
Following the same analysis in the last paragraph, it can be
established that when C is on the left side of S1 , the minimum
is either at point P or at point Q. In
of e2 occurs only when D
summary, through the earlier discussions, we have established
S1 is achieved only when D

that the minimum of e2 for D

is at one of the three boundary points S, Q, or P of S1 . Hence,
the robustness criterion (15) can be rewritten as

2

|C D
 | > 4 2
e2 = |C D|
min

Fig. 7. Examples of joint regions S1 and S2 . (a) Circles C(A, dAC ) and
C(B, dBC ) do not intersect, causing S1 to be joint to S2 . (b) Circles
C(A, dAC ) and C(B, dBC + ) do not intersect, causing S1 to be joint
to S2 .

D{S,Q,P
}
 symmetrical to D
w.r.t. AB
D

D
 | > S .
when |D

(25)

B. Identification of Flip Ambiguities Using Any Pair of

Distance Measurements dAD , dBD , and dCD
The aforementioned analysis analyzed the minimum of e2 =
|C D
 |)2 for D
S1 while using the distance mea(|C D|
surements of sensor node D from neighbors A and B first.
If, instead, the distance measurements of sensor node D from
neighbors A and C are considered first in finding the minimum
|B D
 |)2 , then D
will be symmetric to D
 with
of e2 = (|B D|
respect to AC, and the confined regions S1 and S2 (together
with the intersection points S, P , and Q) will be determined by
the measured distances dAD and dCD and the threshold . The
intersection points S, P , and Q with respect to AC (let them be
S AC , P AC , and QAC ) will be different from their counterparts
with respect to AB (i.e., S AB , P AB , and QAB ), leading to
varying minimum values of e2 . Hence, the minimum e2 for all
three permutations of the sequence A, B, and C should be

2

|Z D
 | > 4 2
e2 = |Z D|
min
XY ,QXY ,P XY }

D{S
 symmetrical to D
w.r.t. XY
D
X,Y,Z{A,B,C},X=Y =Z

D
 | > S .
when |D

(26)

The results of our analysis in this section are summarized in

the following propositions.
Proposition 1: Given fixed positions A, B, and C and for
given measurements dAD , dBD , and dCD satisfying S /2 <
dAD , dBD , dCD R, if there are two disjoint regions S1 and
S2 defined by (12) and (13), then the minimum of e2 in (15)
occurs at S, P or Q, where S, P , and Q are defined right before
(16) and illustrated in Fig. 5.
Proposition 2: For given fixed positions A, B, and C and
for given measurements dAD , dBD , and dCD satisfying S /2 <
dAD , dBD , dCD R, the quadrilateral ABCD satisfying (26)
does not suffer from substantial flip ambiguity.
The aforementioned analysis only considered the situation in
which the two regions S1 and S2 are disjoint. If instead they
are joint, as shown in Fig. 7, the minimum of e2 will always

Fig. 8. More precisely defined two regions S1 and S2 for the possible location
D
 ) of sensor node D using the accurate locations of sensor
estimation pair (D,
nodes A, B, and C and intersensor distance measurements dAD , dBD , and
dCD . Here, S1 and S2 are the intersection of the three rings R1 , R2 , and R3
defined by (27)(29).

and D
 both coincide at the same point on
be zero when D
the common boundary of S1 and S2 . In such situations, the
FCQs will always be detected as not robust. This scenario can
happen more frequently when either the noise in the distance
measurement is large or the chosen threshold  is large and may
cause an increase in false alarms.
C. Identification of Flip Ambiguities When Regions S1
and S2 Are Considered Nonsymmetrical
The robustness criterion (25) formulated in Section VI-A
and D
 to be symmetric to the line AB. This
considers D
assumption is reasonable while considering that the regions
S1 and S2 are formed by the constraints imposed by two
neighboring sensors A and B. Instead, if all three constraints
imposed by the three neighboring sensors A, B, and C are
considered, the possible regions S1 and S2 for the possible
D
 ) will be formed by the intersection of the
location pair (D,
three rings defined by

R1 = (x, y) | ( dAD  )2 (x xA )2 + (y yA )2

(27)
( dAD +  )2

2
2
2
R2 = (x, y) | ( dBD  ) (x xB ) + (y yB )

(28)
( dBD +  )2

2
2
2
R3 = (x, y) | ( dCD  ) (x xC ) + (y yC )

(29)
( dCD +  )2
and are more likely to be nonsymmetrical. An example of
nonsymmetrical regions S1 and S2 is shown in Fig. 8.
Let HC denote the open half-plane with border line AB that
contains C and HC be the complementary half-plane on the

KANNAN et al.: ANALYSIS OF FLIP AMBIGUITIES FOR ROBUST SENSOR NETWORK LOCALIZATION

2065

other side of AB. Without loss of generality, let S1 HC ,

which makes S2 HC . Thus, S1 and S2 can be redefined as
S1 = {(x, y) | (x, y) (R1 R2 R3 HC )}


S2 = (x, y) | (x, y) (R1 R2 R3 HC ) .

(30)
(31)

Note that the true location D lies in only one of the two
regions S1 or S2 , and thus, one of them always exist. Depending on the known locations of A, B, and C, the corresponding distance measurements dAD , dBD , and dCD , and the
threshold , the other region may or may not exist. With these
definitions and analysis, the robustness criterion (25) can be
redefined as

2

|C D
 | > 4 2
e2 = |C D|
min

Fig. 9. (a) Ratio of correct flip detection to all the flip occurrences. (b) Ratio
of missed flip detection to all the flip occurrences.


DS
1 ,D S2

D
 | > S . (32)
when |D
To generalize, when X, Y, Z {A, B, C} and X = Y = Z,
let HZ denote the open half-plane with border line XY that
contains Z, and let HZ be the complementary half-plane on the
other side of XY . Thus, SZ and SZ can be defined as
SZ = {(x, y) | (x, y) (R1 R2 R3 HZ )}


SZ = (x, y) | (x, y) (R1 R2 R3 HZ )

(33)
(34)

and the consideration of minimum e2 for all three permutations

of the sequence A, B, and C will change (32) as follows:

2

|Z D
 | > 4 2
e2 = |Z D|
min

DS
Z ,D SZ
X,Y,Z{A,B,C},X=Y =Z

D
 | > S . (35)
when |D
Finding a closed-form solution for (35) is a complex problem.
Instead, a constrained minimization can be applied to evaluate
(35) for any sensor quadruple ABCD. Since the robustness criterion is more likely to be applied on every sensor quadruple in
a localization algorithm, a robustness criterion without a closedform solution is not suitable for sensor network localization.
VII. N UMERICAL A NALYSIS OF THE
ROBUSTNESS C RITERION
This section tests and compares the performance of different
robustness criteria in detecting the possible flip ambiguities in
the localization of sensor quadruples. In this numerical study,
a pool of |NX |-node sensor networks is created by uniformly
distributing |NX |(= 4) neighbors in a rectangular region of
100 m 100 m with a transmission range of R = 10 m.
Following Assumption 1 in Section V, two sensor nodes X
and Y are said to be neighbors if and only if their intersensor
distance is less than or equal to the transmission range R. The
measured distance between any neighboring sensor node pair is
obtained by blurring the intersensor distance with a Gaussian
noise [36] of zero mean and variance 2 as
dXY = dY X = |XY | + N (0, 2 ).

(36)

Since the distances are always positive, the Gaussian noise is

truncated such that 0 dXY .

Fig. 10. (a) Ratio of correct detection when there is no flip ambiguity to all
the occurrences of estimation without flip ambiguities. (b) Ratio of false alarm
as flipped realization when there is no flip ambiguity to all the occurrences of
estimation without flip ambiguities.

To compare different scenarios, various simulations are done

with different pools of |NX |-sensor networks created by varying the standard deviation of the Gaussian noise in (36) from
0.05 to 0.5 m. In each simulation, 10 000 FCQs are randomly
selected from each pool of |NX |-sensor networks, and localization is done by minimizing the cost function (38). To compare
different scenarios, the same set of simulations are done with
three different robustness criteria:
i) using robustness criterion (26);
ii) using the robustness criterion in [15];
iii) using robustness criterion (35).
The first set of simulations tested the effect of different
thresholds on criteria (25) and (32) in detecting the flip ambiguities while having a fixed noise level of = 0.2 m. A numerical
study in [18] shows that the estimation error will be within a
value of 20% of the transmission range R when there is no
flip ambiguity, but when there is a flip ambiguity, the error
will be much larger than that. Thus, to simplify the detection
of flipped realization, any estimation error > R/5 is treated
as having substantial flip ambiguity. The obtained results are
shown in Figs. 9 and 10. Fig. 9(a) shows the ratio of correct
flip detection to all flipped estimates, and Fig. 9(b) shows the
ratio of missed flip detection to all flipped estimates. It can be
seen from these two figures that the rate of correct flip detection
increases while the rate of missed flip detection decreases with
increasing value of . Fig. 10(a) shows the ratio of correct
detection to all estimates without flip, and Fig. 10(b) shows the
ratio of estimates without flip ambiguities false alarmed to be
flipped estimates to all the estimates without flip ambiguities.
These two figures show that if the threshold is increased, the
false-alarm rate is also increased. This is due to more frequent
joining of regions S1 and S2 with larger thresholds. The effect
of different noise levels () on the correct detection of flipped
estimates or estimates with no flip ambiguity is also studied
with a threshold of  = 3. The results are shown in Fig. 11.
The second set of simulations repeated the previous set of
simulations while removing the dependence on the choice of

2066

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 59, NO. 4, MAY 2010

not be able to localize a significant number of sensor nodes.

However, it has been well established that localization results
will always be poor for sensor nodes that are not uniquely
localizable, and it is better to eliminate these sensor nodes from
the localization process [15], [17].

Fig. 11. Total number of correct detection against when the threshold is set
as  = 3.

Fig. 12. (a) Ratio of correct flip detection to all the flip occurrences. (b) Ratio
of missed flip detection to all the flip occurrences.

Fig. 13. (a) Ratio of correct detection when there is no flip to all the
occurrences of estimation without flip. (b) Ratio of false alarm as flip when
there is no flip to all the occurrences of estimation without flip.

VIII. P ERFORMANCE E NHANCEMENT OF L OCALIZATION

A LGORITHMS U SING THE ROBUSTNESS C RITERION
This section enhances the performance of the localization
algorithms mentioned in Section IV by applying the robustness
criterion (26) in the neighborhood selection of the algorithms.
Our robustness criterion (26) uses only the location and intersensor distance information of the neighbor node and can
thus be used with any valid candidate, i.e., either one-hop or
multihop neighbor, of the neighbor set ND of each node D
to be localized. However, to keep the simulation procedure
simple, only one-hop neighbors are considered to be the valid
candidates of neighbor set ND .
The localization algorithms mentioned in Section IV use
the neighbor set ND to do either trilateration with any three
members of ND or multilateration with all members of ND . In
this paper, the three members of ND for trilateration is selected
using the robustness criterion (26) as follows: The neighbor
|N |
set ND is used to obtain sets C3 D of all possible FCQs with
sensor nodes A, B, C ND and D and to find the most robust
FCQ satisfying

2

|ZD
|
min
e2 = |ZD|
max
XY ,QXY ,P XY }
A,B,CND

D{S
A=B=C
X,Y,Z{A,B,C},X=Y =Z
FCQ ABCD satisfy (26) 
w.r.t. XY
D symmetrical to D

(37)

Fig. 14. Ratio of correct detection against when the threshold is set
as  = 3.

the sequence of points A, B, and C in the analysis by using

criteria (26) and (35) instead of (25) and (32).
Figs. 12 and 13 show the effect of different thresholds ()
on the correct flip detection at a fixed noise level ( = 0.2 m).
Fig. 12(a) and (b) shows that when the dependence on the
choice of the sequence of points A, B, and C is removed, the
rate of correct flip detection is (almost) 100%, and the missed
flip detection is (almost) 0% when the value  3 (i.e., when
the error bound is accurate). Fig. 13(a) and (b) shows that, with
increasing threshold, the false-alarm rate gets higher. This is
again due to more frequent joining of regions S1 and S2 with
larger thresholds. Fig. 14 shows the effect of different noise
levels on the correct detection of flip or no-flip ambiguity
by (26) and (35) at a threshold  of 3. From Figs. 1214, it
can be seen that the robustness criterion (26) makes the correct
decision more accurately than the robustness criterion in [15],
where the assumption of symmetry is creating a considerable
amount of false alarms compared with (35). The drawback of
the criterion (26), as well as the criterion in [15], is that under
conditions of high measurement noise, the algorithms may

to form the robust neighbor set RND to be used in trilateration.

In the case of mutilateration, the robust neighbor set RND
is selected in a particular way as follows: First, the robust
neighbor set RND is formed with the most robust FCQ as
in trilateration. Then, every neighbor node Y ND \ RND is
checked to see whether Y and D form a robust FCQ, together
with every pair of sensor nodes in RND . If they do, then the
robust neighbor set RND is enlarged as RND := RND Y .
The next section give details of the use of the robustness
criterion (26) in sequential localization algorithm and checks
the performance enhancements via simulations. A similar study
was done for a cluster-based localization algorithm, and the performance enhancement was noted to be similar to the sequential
localization algorithm. To restrict the length of this paper, this
algorithm and the simulation results are omitted in this paper.
A. Robustness Criterion in Sequential Localization
Algorithm: In a sequential (or incremental) localization algorithm, every unknown sensor node X checks how many
of its known neighbors Y NX are eligible to be a member
of the robust neighbor set RNX . If they have at least three
members in RNX , they localize themselves, become elevated
anchors, and broadcast their location information together with
their elevated anchor status to all their neighbors Y MX . On
receiving elevated anchor status messages, an unknown sensor
node updates its location-known neighbor set NX . Before an
unknown sensor node tries to localize itself each time, it checks

KANNAN et al.: ANALYSIS OF FLIP AMBIGUITIES FOR ROBUST SENSOR NETWORK LOCALIZATION

and processes any pending elevated anchor status messages. An

unknown sensor node keeps trying to localize itself until either
1) it becomes an elevated anchor, or 2) it does receives no more
elevated anchor messages. When all the unknown sensor nodes
reach this state, the localization algorithm ends. At this point,
there may be no sensor nodes localized due to RNX having less
than three anchors at the initial stage, or there may be one or
more distinct clusters of sensor nodes localized. Irrespective of
the number of localized clusters in the sensor network, all the
localized sensor nodes are localized in the central coordinate
system fixed by the anchor nodes.
Simulation Results: In applications like bushfire surveillance, water-quality surveillance, and precision irrigation, a
large number of tiny sensors are randomly distributed over
a large regular-shaped area and are stationary after being
deployed. In such networks, a sequential algorithm is sufficient, and thus, the performance of both trilateration- and
multilateration-based sequential localization algorithms is evaluated here via simulations, where the robust neighbor set RNX
is chosen using (26).
For a sensor network with a large number of sensor nodes, it
is reasonable to assume the sensor distribution to be uniform,
and such an assumption has widely been used in the area
[5][7], [15], [22], [33]. Changing the sensor distribution
method may only affect the frequency of collinear placement of
sensor nodes. The focus of this paper is on identifying possible
flip ambiguities by neighborhood geometries once the sensor
nodes are placed and eliminating them from the localization
process, and thus, this simulation only considers uniformly
distributed sensor networks to demonstrate the performance
enhancement. In the simulation, 500 different sensor networks
are used, where different random seeds are used to populate
each sensor network with 100 sensor nodes, which are uniformly distributed in a square region of 100 m 100 m.
In these 100 nodes, the first ten are chosen as anchor nodes
and are initialized with random coordinates within the square
region. The measured distance between any neighbor node pair
is modeled using (36), where the standard deviation is chosen
to be 0.2 m. The location of the unknown sensor nodes are
estimated using
 
2
pX = arg min
dXY 
pX pY 
(38)
pX

Y Ni

where pX = (
xX , yX ) denotes any possible location estimate
of sensor node X, dXY is the measured distance between
sensor nodes X and Y , and NX ( MX ) is the set of neighbors
of X whose locations are known or already estimated in the
localization process.
The average node degree is varied in the simulation from 4
to 25 by adjusting the transmission range while keeping the
number of sensor nodes fixed. Fig. 15 shows the relationship
of the average transmission range and the average node degree.
The average mean square error (MSE) in location estimates
is calculated and normalized to the transmission range R as
follows:
MSE =

1
|n|{Vn = }|

500

n=1
|Vn |=0

iVn(xX

x
X )2 + (yX yX )2
|Vn |R2

2067

Fig. 15. Relationship of range versus average node degree.

Fig. 16. Example of trilateration-based sequential localization having flip

ambiguity and removal of such flip ambiguity by the robustness criterion.
Multilateration-based sequential localization has similar performance in the
removal of flip ambiguity. Here, the red line segments show the error offset
between the original and the estimated location of the sensor node. (a) Localization without robustness criterion. (b) Localization with robustness criterion
(26). It is worth noting here that, in this simulation, not many unknown
sensor nodes have enough anchor nodes or elevated anchor nodes in their
neighborhoods, resulting in an early termination of the sequential localization
algorithm. Thus, only a small number of sensor nodes are localized. Since the
total number of the sensor nodes is fixed in the simulations, the number of
neighbors in a neighborhood is restricted by the transmission range, which is
set to be 15 m in this case.

where (
xX , yX ) and (xX , yX ) are the estimated location and
true location of sensor node X, and Vn is the set of sensor nodes
localized in the nth sensor network, whereas |Vn | is the number
of sensor nodes in Vn .
The first set of simulations are done using a fixed threshold
 = 3 in the robustness criterion (26). To compare different
scenarios, trilateration-based localization and multilaterationbased localization are categorized into two groups, and the
neighbors to be used in the location estimate of each group are
chosen using three different methods:
A. trilateration
A.i) any three neighbors of NX ;
A.ii) three most robust neighbors of NX by the criterion
in [15];
A.iii) three most robust neighbors of NX by criterion
(26);
B. multilateration
B.i) all neighbors of NX ;
B.ii) all robust neighbors of NX by the criterion in [15];
B.iii) all robust neighbors of NX by criterion (26).
An efficient flip ambiguity removal by the robustness criterion (26) in trilateration-based sequential localization is shown
in Fig. 16. A numerical study in [18] shows that the location
estimation errors are generally low, and a large variance in
location estimation error occurs only due to flipped realization and the following avalanche effects. The ten-time-lower
average estimation error of A.ii) and A.iii), and B.ii) and
B.iii) compared with A.i) and B.i), respectively, in Fig. 17(a)
and (c), can be interpreted as efficient flip ambiguity removal
by the robustness criteria in the entire trilateration-based and
multilateration-based sensor network localization.

2068

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 59, NO. 4, MAY 2010

Fig. 17. Performance of different neighbor selection methods. (a) and

(b) Trilateration-based sequential localization. (c) and (d) Multilateration-based
sequential localization. Note that the curves (A.ii) and (A.iii) and (B.ii) and
(B.iii) overlap each other and are hence not clearly conspicuous.

Fig. 19. Performance of different thresholds  in (26). (a) and

(b) Trilateration-based sequential localization. (c) and (d) Multilateration-based
sequential localization. Note that the curves  = 3 and  = 4 overlap each
other in (a) and (c) and are hence not clearly conspicuous.

and 4. The results are shown in Fig. 19. As expected,

the results show that the false alarm reduces with decreasing
threshold , increasing the number of localized nodes. It also
shows that missed detection increases with decreasing threshold
, increasing the average estimation error of localized nodes.

Fig. 18. Average localized nodes. (a) Trilateration-based sequential localization. (b) Multilateration-based sequential localization.

As discussed in Section VII, to simplify the flip detection,

any estimation error > R/5 is treated as having substantial
flip ambiguity. Fig. 17(b) and (d) shows the sensor network
realization with possible flip ambiguities detected under this
definition, and it can bee seen that neighbor selection methods
A.i) and B.i) resulted in more flipped realizations in many
sensor network realizations. This is an expected result, since
the numerical study presented in Section VII showed that the
robustness criterion removes all flip ambiguities when  3.
The numerical study also showed that the considerable amount
of false alarms caused by the robustness criterion does not allow
some of the location estimates to happen. This is the reason
for the lower average localized nodes seen in A.ii) and A.iii),
and B.ii) and B.iii) compared with A.i) and B.i), respectively,
in Fig. 18.
It is worth noting here that the number of localized nodes
using both trilateration and multilateration are the same and that
the difference is in the estimation errors created by them. When
the FCQs with possible flip ambiguity are not identified and
removed from the localization process, multilateration helps remove some but not all flip ambiguities, creating less estimation
error compared with its counterpart, i.e., trilateration. However,
when the FCQs with possible flip ambiguity are identified and
removed from the localization process, the difference in the
estimation errors between the two methods are marginal, which
is much smaller than the previous case.
The relationship of false alarm and missed detection with
respect to the threshold  is seen in Section VII. To see the effect
of the varying threshold  in the sensor network localizations,
a second set of simulations are done with variable threshold
 in the robustness criterion (26), where  is varied between

IX. T HRESHOLD E FFECTS AND AN I TERATIVE

M ULTITHRESHOLD L OCALIZATION M ETHOD
A. Algorithm
As seen in Section VII, the robustness criterion with a larger
threshold  may cause more false alarms. In such situations,
more sensor nodes with three or more members in their neighbor set cannot be localized by the localization algorithm using
our criterion. On the other hand, a smaller threshold  may
cause more missed flip detection. In such situations, there
may be more sensor nodes with faulty localization results.
To prevent both of these situations, an iterative localization
algorithm is presented in this section that can be applied to any
of the localization algorithms presented in the last section.
This iterative localization method starts with a large threshold 1 to obtain an initial location estimate. In the subsequent
iterations, the threshold i1 (for i = 2, 3, . . .) is gradually
reduced by a predefined step size  as i = i1  before using it in the robustness criterion. This reduction of i
in each iteration causes a reduction in the number of false
alarms and, in return, increases the number of localized nodes.
This algorithm can be iterated until the threshold i reaches
zero or some predefined minimum value min , which can be
determined based on the required accuracy level of the location
estimates.
After each iteration of the localization algorithm, the status
of the sensor nodes localized in that iteration will be promoted
to elevated anchor nodes. The subsequent iterations will not
reestimate the locations of these elevated anchor nodes. Only
the unknown sensor nodes with a robust neighbor set RNX
satisfying the robustness criterion with the reduced threshold
i will be localized in the corresponding iteration step i. The
reduction of threshold i at each subsequent iteration is likely
to cause more missed detection, in return causing more flip
ambiguities than in the previous iterations. Thus, the estimation

KANNAN et al.: ANALYSIS OF FLIP AMBIGUITIES FOR ROBUST SENSOR NETWORK LOCALIZATION

2069

multilateration-based iterative localization algorithms localized

the same number of sensor nodes. Hence, only the results of
the multilateration-based iterative localization algorithm are
presented in Fig. 20(c) and (d).
It can be seen from the results shown in Fig. 20 that our iterative localization algorithm significantly improves the number of
localized sensor nodes (within an acceptable estimation errors)
compared with the number of localized sensor nodes presented
in Fig. 18.
X. C ONCLUSION

Fig. 20. Performance of the iterative process of the sequential localization. (a), (c), and (d) Trilateration-based sequential localization. (b), (c), and
(d) Multilateration-based sequential localization. Note that the curves  = 0
and i = 3 > 2 > 1 > 0 are an exact match of each other in (c)
and are hence inconspicuous.

error is likely to increase more in each subsequent iteration than

in the previous iterations.
The iterative algorithm is summarized as follows, where the
number of iterations n, n = min , and 1 =  are predefined;
the vertex set Va represents the set of anchor nodes, and the
vertex set V represents the complete set of sensors of the
underlying network graph.
Algorithm 1
Iterative
Multithreshold
Localization
Algorithm
1: 1 
2:  ( min /n)
3: V0 Va
4: for i = 1 to n do
5: Apply localization algorithm with robustness criterion
and threshold i , to obtain the location estimates of the
localizable sensors in V \ i1
j=0 Vj .
6: Label the set of newly localized sensors as Vi .
7: i+1 = i 
8: end for
Using the iterative method described earlier, we obtain better
localization accuracy of the sensor nodes than the localization
accuracy that we would have obtained using a single threshold
value.
B. Simulation Results
The performance of the iterative sequential localization algorithm is presented in this section. Even though the iterative
cluster localization algorithm produces a similar result, due
to page limitations, the corresponding simulation results are
not presented in this paper. The simulations are done in four
iterative steps by starting with a threshold value = 3 and
reducing it with a step size of  = until the threshold 
reaches zero. Each iteration of the sequential localization algorithm used the reduced threshold  in the robustness criterion
(26), while taking all localized nodes of previous iterations as
elevated anchors. The results of each iterative step are shown
in Fig. 20. As noted in Section VIII-A2, both trilateration- and

In this paper, we have analyzed flip ambiguity problems and

their implications in planar sensor network localization by a
formal geometric analysis. We have quantified the likelihood of
flip ambiguities in arbitrary sensor neighborhood geometries.
We also have developed a robustness criterion that can identify
the likelihood of flip ambiguity in FCQs, which are quadruples
of sensor nodes, all of which are neighbors of each other, via
a formal optimality analysis. This robustness criterion is used
in localization algorithms to eliminate sensor nodes from any
neighborhood that contribute to flip ambiguities. The simulation results demonstrate that the benefits of using our proposed
robustness criterion are significant, and the criterion performs
well in identifying flip ambiguities while causing a very small
number of false alarms. The simulation results also show that
the use of the robustness criterion in an iterative multithreshold
localization method improves the number of localized nodes
in a sensor network while keeping the accuracy of estimation
errors at acceptable levels.
R EFERENCES
[1] S. Guolin, C. Jie, G. Wei, and K. Liu, Signal processing techniques
in network-aided positioning: A survey of state-of-the-art positioning designs, IEEE Signal Process. Mag., vol. 22, no. 4, pp. 1223,
Jul. 2005.
[2] F. Gustafsson and F. Gunnarsson, Mobile positioning using wireless
networks: Possibilities and fundamental limitations based on available
wireless network measurements, IEEE Signal Process. Mag., vol. 22,
no. 4, pp. 4153, Jul. 2005.
[3] A. Sayed, A. Tarighat, and N. Khajehnouri, Network-based wireless
location: Challenges faced in developing techniques for accurate wireless
location information, IEEE Signal Process. Mag., vol. 22, no. 4, pp. 24
40, Jul. 2005.
[4] G. Mao, B. Fidan, and B. Anderson, Wireless sensor network localization
techniques, Comput. Netw., vol. 51, no. 10, pp. 25292553, Jul. 2007.
[5] G. Mao, B. Fidan, and B. Anderson, Sensor network localization, in
Sensor Network and Configuration: Fundamentals, Techniques, Platforms
and Experiments, N. P. Mahalik, Ed. New York: Springer-Verlag, 2006,
pp. 281316.
[6] Y. Shang, W. Ruml, Y. Zhang, and M. Fromherz, Localization from
connectivity in sensor networks, IEEE Trans. Parallel Distrib. Syst.,
vol. 15, no. 11, pp. 961974, Nov. 2004.
[7] P. Biswas and Y. Ye, Semidefinite programming for ad hoc wireless
sensor network localization, in Proc. Inf. Process. Sens. Netw., 2004,
pp. 4654.
[8] A. Kannan, G. Mao, and B. Vucetic, Simulated annealing based wireless
sensor network localization with flip ambiguity mitigation, in Proc. IEEE
Veh. Technol. Conf., 2006, vol. 2, pp. 10221026.
[9] K. Langendoen and N. Reijers, Distributed localization in wireless sensor
networks: A quantitative comparison, Comput. Netw., vol. 43, no. 4,
pp. 499518, Nov. 2003.
[10] V. Fox, J. Hightower, L. Lin, D. Schulz, and G. Borriello, Bayesian
filtering for location estimation, IEEE Pervasive Comput., vol. 2, no. 3,
pp. 2433, Jul.Sep. 2003.
[11] A. Ihler, I. Fisher, R. Moses, and A. Willsky, Nonparametric belief
propagation for self-localization of sensor networks, IEEE J. Sel. Areas
Commun., vol. 23, no. 4, pp. 809819, Apr. 2005.

2070

IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 59, NO. 4, MAY 2010

[12] J. Aspnes, T. Eren, D. Goldenberg, A. Morse, W. Whiteley, Y. Yang,

B. Anderson, and P. Belhumeur, A theory of network localization, IEEE
Trans. Mobile Comput., vol. 5, no. 12, pp. 16631678, Dec. 2006.
[13] R. Connelly, Generic global rigidity, Discrete Comput. Geometry,
vol. 33, no. 4, pp. 549563, Apr. 2005.
[14] T. Eren, D. Goldenberg, W. Whiteley, Y. Yang, A. Morse, B. Anderson,
and P. Belhuneur, Rigidity, computation, and randomization in network
localization, in Proc. IEEE INFOCOM, 2004, vol. 4, pp. 26732684.
[15] D. Moore, J. Leonard, D. Rus, and S. Teller, Robust distributed network
localization with noisy range measurements, in Proc. 2nd ACM Conf.
Embed. Netw. Sens. Syst., 2004, pp. 5061.
[16] B. Hendrickson, Conditions for unique graph realizations, SIAM J.
Comput., vol. 21, no. 1, pp. 6584, Feb. 1992.
[17] D. Goldenberg, A. Krishnamurthy, W. Maness, Y. Yang, and A. Young,
Network localization in partially localizable networks, in Proc. IEEE
INFOCOM, 2005, pp. 313326.
[18] A. Kannan, B. Fidan, and G. Mao, Robust distributed sensor network localization based on analysis of flip ambiguities, in Proc. IEEE
GLOBECOM, 2008, pp. 16.
[19] S. Lederer, Y. Wang, and J. Gao, Connectivity-based localization of large
scale sensor networks with complex shape, in Proc. IEEE INFOCOM,
2008, pp. 789797.
[20] Y. Wang, S. Lederer, and J. Gao, Connectivity-based sensor network localization with incremental delaunay refinement method, in Proc. IEEE
INFOCOM, 2009, pp. 24012409.
[21] Z. Yang, Y. Liu, and X. Li, Beyond trilateration: On the localizability
of wireless ad-hoc networks, in Proc. IEEE INFOCOM, 2009, pp. 2392
2400.
[22] D. Niculescu and B. Nath, Dv based positioning in ad hoc networks,
Telecommun. Syst., vol. 22, no. 14, pp. 267280, Jan. 2003.
[23] A. Savvides, C. C. Han, and M. B. Srivastava, Dynamic fine-grained
localization in ad-hoc networks of sensors, in Proc. ACM SigMobile,
2001, pp. 166179.
[24] S. Capkun, M. Hamdi, and J. Hubaux, Gps-free positioning in mobile
ad-hoc networks, in Proc. 34th Hawaii Int. Conf. Syst. Sci., 2001,
pp. 34813490.
[25] F. Sittile and M. Spirito, Robust localization for wireless sensor
networks, in Proc. IEEE SECON, 2008, pp. 4654.
[26] A. Kannan, B. Fidan, G. Mao, and B. Anderson, Analysis of flip ambiguities in distributed network localization, in Proc. Inf., Decision Control,
2007, pp. 193198.
[27] J. Fang, M. Cao, A. Morse, and B. Anderson, Localization of sensor
networks using sweeps, in Proc. IEEE Conf. Decision Control, 2006,
pp. 46454650.
[28] D. Goldenberg, P. Bihler, M. Cao, J. Fang, B. Anderson, A. Morse,
and Y. Yang, Localization in sparse networks using sweeps, in Proc.
ACM/IEEE MobiCom, 2006, pp. 110121.
[29] B. Anderson, P. Belhumeur, T. Eren, D. Goldenberg, A. Morse,
W. Whiteley, and R. Yang, Graphical properties of easily localizable
sensor networks, Wireless Netw., vol. 15, no. 2, pp. 177191, Feb. 2009.
[30] C. Savarese and J. Rabaey, Robust positioning algorithms for distributed
ad-hoc wireless sensor networks, in Proc. General Track: USENIX Annu.
Techn. Conf., 2002, pp. 317327.
[31] C. Savarese, J. Rabaey, and J. Beutel, Location in distributed ad-hoc
wireless sensor networks, in Proc. ICASSP, 2001, pp. 20372040.
[32] L. Meertens and S. Fitzpatrick, The distributed construction of a global
coordinate system in a network of static computational nodes rom
inter-node distances, Kestrel Institute, Palo Alto, CA, Formal Rep.
KES.U.04.04, 2004.
[33] X. Ji and H. Zha, Sensor positioning in wireless ad-hoc sensor networks
using multidimensional scaling, in Proc. IEEE INFOCOM, 2004, vol. 4,
pp. 26522661.
[34] Y. Shang and W. Ruml, Improved MDS-based localization, in Proc.
IEEE INFOCOM, 2004, vol. 4, pp. 26402651.
[35] Y. Kwon, K. Mechitov, S. Sundresh, W. Kim, and G. Agha, Resilient
localization for sensor networks in outdoor environments, in Proc. IEEE
ICDCS, 2005, pp. 643652.
[36] N. Patwari, A. Hero, III, M. Perkins, S. Neiyer, and R. ODea, Relative
location estimation in wireless sensor networks, IEEE Trans. Signal
Process., vol. 51, no. 8, pp. 21372148, Aug. 2003.

Anushiya A. Kannan (M08) received the B.Sc. and

B.E. degrees, in 1994 and in 1995, respectively, from
the University of Sydney, Sydney, Australia, where
she is currently working toward the Ph.D. degree in
wireless sensor networks.
She has been with TriTech Microelectronics,
Singapore; Adaptive Broadband, Cambridge, U.K.;
and Cellonics, Singapore. She is currently a Researcher with the Biomedical Systems Laboratory,
University of New South Wales, Sydney. Her research interests include wireless sensor networks,
wireless communications, biomedical engineering, and integrated circuit
design.

Bars Fidan (M03) received the B.S. degrees

in electrical engineering and mathematics from
Middle East Technical University, Ankara, Turkey, in
1996, the M.S. degree in electrical engineering from
Bilkent University, Ankara, in 1998, and the Ph.D.
degree in electrical engineering from the University
of Southern California, Los Angeles, in 2003.
He was with the University of Southern California
in 2004 as a Postdoctoral Research Fellow and with
the National ICT Australia and the Research School
of Information Sciences and Engineering, Australian
National University, during 20052009 as a Researcher/Senior Researcher. He
is currently an Assistant Professor with the Department of Mechanical and
Mechatronics Engineering, University of Waterloo, Waterloo, ON, Canada. His
research interests include autonomous multiagent dynamical systems, sensor
networks, cooperative localization, adaptive and nonlinear control, switching
and hybrid systems, mechatronics, and various control applications, including
high-performance and hypersonic flight control, semiconductor manufacturing
process control, and disk-drive servo systems.

Guoqiang Mao (SM08) received the Bachelors degree in electrical engineering from Hubei Polytechnic University, China, in 1995, the Masters degree
in engineering from South East University, Nanjing,
China, in 1998, and the Ph.D. degree in telecommunications engineering from Edith Cowan University,
Perth, Australia, in 2002.
In December 2002, he joined the School of Electrical and Information Engineering, University of
Sydney, Sydney, Australia, where he is currently
a Senior Lecturer. He has published more than
70 papers in journals and refereed conference proceedings. His research interests include wireless localization techniques, wireless multihop networks,
graph theory and its application in networking, and network performance
analysis.
Dr. Mao has served as a program committee member at a number of
international conferences and was the publicity co-chair of the 2007 ACM
SenSys and the publicity co-chair of the 2010 IEEE Wireless Communications
and Networking Conference.